AitopkitMicrosoft is going bigger on its own silicon. At the OCP Global Summit on October 12, the company announced Maia 300, its third-generation custom AI inference accelerator — a chiplet-based chip built on TSMC 2nm/3nm process technology with eight HBM4 memory stacks, packed 96 chips to a rack.
The headline claims are about efficiency at data-center scale: Microsoft says Maia 300 delivers ~40% more tokens per megawatt than leading GPU-based systems and up to 7x the rack-scale throughput of its predecessor, Maia 200. If those numbers hold in production, it would mean meaningfully cheaper inference for the same compute — exactly the cost lever that matters as AI workloads keep exploding.
The strategy goes well beyond one chip generation. Starting in 2027, Microsoft plans to scale Azure Maia to multi-gigawatt deployments, positioning its in-house silicon as the primary inference platform for Copilot, GitHub Copilot, Microsoft Teams, and OpenAI models served on Azure. CEO Satya Nadella called the effort "our biggest bet yet on our own silicon" — a direct challenge to Nvidia's dominance in AI compute.
For users, the takeaway is second-order but important: cheaper, more efficient inference chips eventually show up as cheaper API pricing and faster AI features. Microsoft's bet is that owning the hardware underneath Azure's AI services is how it keeps that cost curve bending down.